Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AT-RAG: An Adaptive RAG Model Enhancing Query Efficiency with Topic Filtering and Iterative Reasoning

Mohammad Reza Rezaei, Maziar Hafezi, Amit Satpathy, Lovell Hodge, Ebrahim Pourjafari

TL;DR

The integration of topic filtering and iterative reasoning enables the proposed AT-RAG model to handle intricate queries efficiently, which makes it suitable for applications that require nuanced information retrieval and decision-making.

Abstract

Recent advancements in QA with LLM, like GPT-4, have shown limitations in handling complex multi-hop queries. We propose AT-RAG, a novel multistep RAG incorporating topic modeling for efficient document retrieval and reasoning. Using BERTopic, our model dynamically assigns topics to queries, improving retrieval accuracy and efficiency. We evaluated AT-RAG on multihop benchmark datasets QA and a medical case study QA. Results show significant improvements in correctness, completeness, and relevance compared to existing methods. AT-RAG reduces retrieval time while maintaining high precision, making it suitable for general tasks QA and complex domain-specific challenges such as medical QA. The integration of topic filtering and iterative reasoning enables our model to handle intricate queries efficiently, which makes it suitable for applications that require nuanced information retrieval and decision-making.

AT-RAG: An Adaptive RAG Model Enhancing Query Efficiency with Topic Filtering and Iterative Reasoning

TL;DR

The integration of topic filtering and iterative reasoning enables the proposed AT-RAG model to handle intricate queries efficiently, which makes it suitable for applications that require nuanced information retrieval and decision-making.

Abstract

Recent advancements in QA with LLM, like GPT-4, have shown limitations in handling complex multi-hop queries. We propose AT-RAG, a novel multistep RAG incorporating topic modeling for efficient document retrieval and reasoning. Using BERTopic, our model dynamically assigns topics to queries, improving retrieval accuracy and efficiency. We evaluated AT-RAG on multihop benchmark datasets QA and a medical case study QA. Results show significant improvements in correctness, completeness, and relevance compared to existing methods. AT-RAG reduces retrieval time while maintaining high precision, making it suitable for general tasks QA and complex domain-specific challenges such as medical QA. The integration of topic filtering and iterative reasoning enables our model to handle intricate queries efficiently, which makes it suitable for applications that require nuanced information retrieval and decision-making.

Paper Structure

This paper contains 26 sections, 9 equations, 4 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Question Decomposition
  4. Chain-of-Thought with Iterative Retrieval
  5. Adaptive Retrieval
  6. Method
  7. Background
  8. One Step RAG for QA
  9. Multi-Step RAG for QA
  10. Topic Filtering for RAG
  11. AT-RAG Model
  12. Topic Assignment Model
  13. Analysis of Topic Distribution Across Datasets
  14. Data Ingestion for Vector Database
  15. Answer Grader
  16. ...and 11 more sections

Figures (4)

  • Figure 1: Comparison of the average overall score across multiple datasets for different models (One Step , Adaptive , with GPT40). Error bars depict the standard deviations for each model. An ANOVA test st1989analysis (with p<0.05) reveals a statistically significant difference between the and Adaptive , denoted by an asterisk (*). For further details, refer to Table \ref{['tab:main']}
  • Figure 2: The answering pipeline leverages a topic generator to streamline document retrieval. It iteratively generates reasoning steps through a generator, guiding the formulation of answers. This process alternates between retrieval and reasoning until a predefined maximum number of iterations (N) is reached or the answer passes quality checks by grader nodes.
  • Figure 3: Normalized Topic Distribution Using TopicBERT Across Multi-Hop QA Datasets. The bar plot displays the relative density (proportion of total documents) for each topic, highlighting thematic diversity within each dataset. This visualization emphasizes how topic assignment addresses dataset bias, influencing the retrieval process in tasks.
  • Figure 4: (A) A comparison between our and the One Step in answering time-based questions on the medical records of six patients evaluated by GPT-4. (B) The average scores of our proposed approach and the One Step across the six cases. ** indicates a statistically significant difference between the two bars, with p < 0.02 as determined by ANOVA test st1989analysis.